2892

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 48, NO. 10, OCTOBER 2000

Spatial Averaging of TimeFrequency Distributions for Signal Recovery in Uniform Linear Arrays
Yimin Zhang, Member, IEEE, and Moeness G. Amin, Senior Member, IEEE
AbstractThis paper presents a new approach based on spatial timefrequency averaging for separating signals received by a uniform linear antenna array. In this approach, spatial averaging of the timefrequency distributions (TFDs) of the sensor data is performed at multiple timefrequency points. This averaging restores the diagonal structure of the source TFD matrix necessary for source separation. With spatial averaging, crossterms move from their off-diagonal positions in the source TFD matrix to become part of the matrix diagonal entries. It is shown that the proposed approach yields improved performance over the case when no spatial averaging is performed. Further, we demonstrate that in the context of source separation, the spatially averaged WignerVille distribution outperforms the combined spatialtimefrequency averaged distributions, such as the one obtained by using the ChoiWilliams distribution. Simulation examples involving the separation of two sources with close AM and FM modulations are presented. Index TermsArray signal processing, spatial timefrequency distribution, subspace analysis, timefrequency distribution, time frequency MUSIC.

I. INTRODUCTION ECENTLY, timefrequency distributions (TFDs) have been employed for direction finding and blind source separation problems in sensor array processing [1][5]. The spatial timefrequency distributions (STFDs) were introduced in [1] and represented by a spatial matrix whose elements are the auto- and cross-timefrequency distributions of the data received at the different array sensors. STFD techniques are most appropriate to handle sources of nonstationary waveforms that are localized in the timefrequency domain. The robustness of the subspace estimates using STFD matrices is analyzed in [17] and shown to have an advantage over those obtained from the covariance matrices. The application of STFDs to separating sources with distinct timefrequency signatures is presented in [2]. In this reference, it is shown that the source TFD matrix, whose elements are the auto- and cross-TFD of the source signals, and the sensor data STFD have the same relationship as the one between the source and the data correlation matrices. This relationship is defined by the mixing or the array manifold matrix. The steps applied in blind source separation based on second-order statistics
Manuscript received March 30, 1999; revised May 12, 2000. This work was supported by Office of Naval Research under Grant N00014-98-1-0176. The associate editor coordinating the review of this paper and approving it for publication was Dr. Paulo J. S. G. Ferreira. The authors are with the Department of Electrical and Computer Engineering, Villanova University, Villanova, PA 19085 USA (e-mail: yimin@ieee.org; moeness@ece.vill.edu). Publisher Item Identifier S 1053-587X(00)07637-6.

(the SOBI technique) outlined in [9] could therefore be used in the timefrequency formulation of the problem. The general theory of solving blind source separation problems using spatial arbitrary joint-variable distributions, including those of time and frequency, is given in [3]. In [4], the two arbitrary variables are chosen as the time-lag and frequency-lag, and the source separation was performed using spatial ambiguity functions. The use of STFDs for direction finding is discussed in [5] and [18], where the timefrequency MUSIC and the timefrequency maximum likelihood techniques are proposed. Although blind source separations based on TFD outperform the SOBI method for nonstationary signals, the fundamental problem with the bilinear timefrequency approach remains the need for the incorporation of STFD matrices computed only at the source autoterm points. Crossterms impede performance, as they reside on the off-diagonal elements of the source TFD matrix, and as such, violate its diagonal structure necessary for source separation. Identification of autoterm regions are often difficult for a large class of multicomponent nonstationary signals, and even if properly identified, due to the complexity of the impinging signal timefrequency signatures and the use of finite data records, autoterm regions cannot be entirely free from crossterm mainlobe or/and sidelobe contamination. In this paper, we discuss the role of TFD crossterms and demonstrate the effect of spatial averaging on STFDs. By utilizing the Vandermonde structure of the array manifold matrix and performing spatial averaging on the spatial timefrequency distribution matrices, we set the off-diagonal elements of the corresponding source TFD matrix to zero. This is achieved by moving the crossterms from their off-diagonal positions to join the autoterms as diagonal entries of the source TFD matrix at one timefrequency point. In this respect, the performance of the source separation technique becomes much less dependent on the selection of the timefrequency points at which the STFD matrices are computed. It is shown that the spatially averaged STFDs outperform the case where no spatial averaging is performed, even when only autoterm points are involved in both cases. Spatial averaging is a simple and well-known technique in conventional array processing [6]. It employs additional array sensors to reduce cross-correlation in coherent and correlated signal environments, thereby permitting proper angle-of-arrival (AOA) estimations and source separations. In this paper, we show that spatial averaging plays a key role in the underlying TFD-based source separation problem, and its application leads to matrix diagonalization and crossterm mitigation. Spatial averaging gives robustness to timefrequency point selections and yields improved performance over other TFD-based techniques,

1053587X/00$10.00 2000 IEEE

ZHANG AND AMIN: SPATIAL AVERAGING OF TIMEFREQUENCY DISTRIBUTIONS FOR SIGNAL RECOVERY

2893

specifically for sources whose signatures are closely separated in the timefrequency domain. The restoration of the diagonal structure of the source TFD is only part of the problem. Source separation using spatially averaged TFD evaluated at a single timefrequency point can still lead to noisy and nonunique results. Since the power distribution of the signals impinging on the array varies over the timefrequency plane, then different timefrequency points may exhibit different SNRs. The main two advantages of incorporating several spatially averaged TFD matrices evaluated at different timefrequency points into a joint-diagonalization scheme are to avoid having degenerate eigenvalues and to reduce the possibility of choosing a point with high noise contamination. It is noted that unlike the method in [2], the proposed approach requires the information on the array manifold and is sensitive to the calibration error. In this case, conventional AOA estimation methods, such as the maximum likelihood [11], matrix pencil [12], MUSIC [13], root MUSIC [14], [15], and ESPRIT [16] techniques can also be used to estimate the mixing matrix and further to separate the source signals. The proposed approach not only requires no angular search but also enjoys the discriminatory property of TFD-based array processing, where fewer sources can be considered by only selecting their respective timefrequency signatures [17][19]. This paper is organized as follows. In Section II, the source separation approach based on spatial timefrequency distribution is briefly summarized. In Section III, we introduce the spatially averaged timefrequency distributions and discuss the difference between spatial averaging and kernel methods in handling the crossterm problem. Simulation results demonstrating the usefulness of the proposed technique are given in Section IV. II. SOURCE SEPARATION BASED ON SPATIAL TIMEFREQUENCY DISTRIBUTIONS A. Spatial TimeFrequency Distributions for an -eleIn many practical situations, the data vector ment array follows an instantaneous mixture model and is given by (1) is the data snapshot where vector at time , and the superscript denotes transpose. The contains source signals vector is the additive noise vector. This at the same time, and model is commonly used in the field of narrowband array contains processing. The vector the noise-free array output. The mixing matrix is the transfer function between the source signals and the data at the array is full column sensors. We assume that the mixing matrix rank. is assumed to be a deterministic The source signal vector signal vector with correlation matrix (2)

where superscript denotes the conjugate transpose of a madiag trix or a vector. In [2], it is assumed that , where diag is the diagonal matrix formed with the elements of its vector valued argument, and denotes the cor. This assumption implies that the components relation of , are mutually uncorrelated. However, in our proposed method, this assumption is no longer necessary. is modeled as a stationary, temporally The additive noise white, zero-mean complex random process independent of the to be spasource signals. For simplicity, we also require tially white, i.e., (3) is the Kronecker delta, and denotes the identity where matrix. Since the signal power and the signal ordering are indeterminable in source separations [3], we simplify the problem by treating the source signals as if they have unit power. Accordingly and (4) The discrete-time form of Cohens class of TFD for signal is given by [7]

(5) where and represent the time index and the frequency index, characterizes the TFD and is a respectively. The kernel function of both the time and lag variables. The cross-TFD of and is defined by [7] two signals

(6) One possible definition of spatial timefrequency distribution (STFD) is given in [2] and incorporates both equations (5) and (6)

(7) for . where It is shown in the next section that other forms of STFD can be more useful in the context of source separation. Under the linear data model of (1), and assuming a noise-free environment, the STFD matrix in (7) takes the following simple structure: (8) is the signal TFD matrix whose entries are the where auto- and cross-TFDs of the sources. Equation(8) is similar to the formula that is commonly used in conventional blind source

2894

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 48, NO. 10, OCTOBER 2000

separation and direction-of-arrival (DOA) estimation problems [8], [9], relating the signal correlation matrix to the data is a full-rank matrix, spatial correlation matrix. If the two subspaces spanned by the principle eigenvectors of and the columns of become identical. In this case, direction-finding techniques based on eigenstructures can be is diagonal, i.e., the signal cross-TFDs at applied. If are zeros, the mixture matrix the timefrequency point and the signal waveforms can be recovered using blind source separation methods [1], [2]. B. Source Separation Based on Spatial TimeFrequency Distributions The source separation algorithm based on spatial timefrequency distributions is an essential part of the proposed method. The algorithm is given in details in [2] and is summarized below. of the The first step is the whitening of the signal part observation. This is achieved by applying a whitening matrix to , i.e., an matrix satisfying

forms STFD matrices at a set timefrequency autoterm points. The unitary of preferable is then obtained as the joint diagonalizer of the set matrix . III. SPATIAL AVERAGING TIMEFREQUENCY DISTRIBUTIONS A. Spatial Averaging Methods The spatial averaging method was introduced by Pillai [6] to restore the full-rank property of the signal correlation matrix in the presence of coherent signals. Unlike other spatial smoothing methods [20][23], which only restore the full-rank property of the mixing matrix when the impinging signals are coherent, the spatial averaging method enforces the diagonal structure of the signal correlation matrix. This diagonal matrix property is essential to perform source separation, as previously discussed. Here, we present the role of spatial averaging in the context of TFD analysis and propose signal separation using joint diagonalization based on spatial averaging of spatial TFD matrices. The basic idea of spatial averaging is to use subarrays of a uniform linear array to obtain an averaged correlation matrix or, in the underlying problem, an averaged STFD matrix, with the off-diagonal elements set to zero. Without loss of generality, we consider a simple example of , i.e., there are only two sources and . The result . is generally true for sources and sensors, as long as By ignoring the effect of noise, the received signal at the th is represented by array sensor (14) are the spatial radian frewhere are the angles-of-arrival, is the RF wavelength, quencies, and is the distance between the zeroth and the th array senand , assuming uniform linear sors. The cross-TFD of array, is

(9) is an as (10) where superscript process denotes pseudo-inverse. The whitened still obeys a linear model (11) By pre- and post-multiplying the STFD matrices , we obtain by (12) which is, in essence, the STFD of the whitened data vector and From the definitions of . (13) is diagonal, then any Equation (13) shows that if whitened data STFD matrix is diagonal in the basis of the columns of the matrix , and the eigenvalues of are the diagonal entries of . An estimate of the may be obtained as a signal subspace of a unitary matrix whitened STFD matrix evaluated at a timefrequency point corresponding to the signal autoterm. The source signals can , and the mixing matrix then be estimated as is estimated by . Although the unitary matrix can be obtained from a single timefrequency point, STFDs corresponding to different points should be incorporated to reduce the possibility of having degenerate eigenvalues and, subsequently, nonunique solutions. The joint-diagonalization (JD) scheme can be used to incorporate multiple timefrequency points [2]. This scheme unitary matrix , and matrix can be written

(15) Due to the presence of the cross-terms [second term in each does not provide bracket in (15)], the TFD matrix the proper information to carry out source separations. and , The auto- and cross-TFD of the data is

(16) as the reference sensor where we used the sensor receiving . Denote and set and . The values of and are generally complex. If and

ZHANG AND AMIN: SPATIAL AVERAGING OF TIMEFREQUENCY DISTRIBUTIONS FOR SIGNAL RECOVERY

2895

are real, then the Hermitian Toeplitz spatial timefrequency matrix

. . .

. . .

..

. . . (17)
Fig. 1. Array configuration for spatial averaging.

, , , generated from the cross-TFDs between the data samples at the reference sensor and those at other sensors of the array can be expressed as [24] (18) where is a Vandermonde matrix, and

is Hermitian and Toeplitz. This matrix is referred to as the spatially averaged TFD (SATFD) matrix. Similar to the real TFD case, in the noise-free environment, the SATFD matrix in (23) can be expressed as (24) where

diag

(19) diag (25) are zero, whereas the The off-diagonal elements of matrix diagonal entries are now made up of both autoterms and crossterms of the impinging source signals. By enforcing the , spatial diagonal structure of the source TFD matrix averaging of the Hermitian Toeplitz STFD matrices extends the validity of the TFD-based signal separation in the presence of cross-TFDs. The steps for source separation used in [2] and summarized instead of in Section II can be applied to the SATFD . With spatial averaging, the incorporation the STFD of STFDs at only autoterm points into the joint-diagonalization scheme is no longer crucial to achieve good performance. B. Comparison between Spatial Averaging and Kernel Methods (21) There are two sources of crossterms in the underlying source separation problem. The first type are the crossterms that are the results of the interactions between the components of the same source signal. Whether we use the STFD defined in (7) or in (17), those crossterms are not harmful to the blind source separation problem since they always reside, along with the autoterms, on the main diagonal of the source TFD matrix. The other type of crossterms are those generated from the interactions between two signal components belonging to two different sources. These crossterms are associated with cross-TFDs of the source signals, and at any given timefrequency point, they constitute the off-diagonal entries of the source TFD matrices. The crossterms generated from the data cross-TFDs violate the basic assumption in the problem of source separation regarding the diagonal structure of the source TFD matrix. We must therefore select the timefrequency points that belong to autoterm regions where crossterm contributions are at minimum. However, the selection of autoterm points is often difficult in the absence of a priori information of the source signals, specifically for low SNR or when the signals have highly overlapping timefrequency signatures. The later case can be encountered in radar echoes and acoustic signal processing.

is the corresponding source TFD matrix. Note that has a different structure from that of the STFD matrix defined in (7) and was used in [2] for blind source separation. Clearly, here is diagonal, (18) has the same form as (8), but point corresponds to a crossterm. even if the selected In the case of complex signal waveforms, the realness and can be restored by spatial the diagonal structure of array sensors symmetrically about the averaging. We add reference point, as shown in Fig. 1. The received signal at th sensor of the new set is (20) The new cross-TFD of and is

The spatial averaging of (16) and (21) is given by

(22) where Re Re Since the terms in the brackets in (21) are all real, the matrix formed from the TFDs (22)

. . .

. . .

..

. . . (23)

2896

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 48, NO. 10, OCTOBER 2000

The use of smoothing timefrequency kernel for crossterm reduction is a candidate solution of the above problem. The main function of this kernel in the context of source separation is to prevent the selection and incorporation of crossterm points in the joint-diagonalization scheme, as well as to reduce the contribution of crossterms at selected autoterm points. In essence, the fundamental role of the timefrequency kernel is to make the source TFD matrices as close to a diagonal structure as possible. The timefrequency kernel can be applied to both forms of STFDs in (7) and (23). It is noteworthy that the smoothing kernel does not distinguish between the aforementioned two types of crossterms, and accordingly, it reduces all entries of the source TFD matrix, including the diagonal elements. The problem with the smoothing kernel is fourfold. First, for sources with closely separated timefrequency signatures, the effectiveness of the smoothing kernel in reducing crossterms is highly impaired. Second, reduction of crossterms depends on their timefrequency locations, especially when fixed shape kernels are used. For example, timefrequency kernels satisfying the marginal properties are not suitable for removing the crossterms which lie on the time-lag and frequency-lag axes in the ambiguity domain. Third, depending on the employed timefrequency kernel, part or all of the crossterms may be displaced to mount on the selected autoterm points. The situation can make the source TFD matrix further deviate from a diagonal structure and cause performance deterioration from the case when no smoothing is applied. We refer to this undesired property as the smoothing problem. Fourth, since source separation is often performed incorporating a finite number of data samples, the intrusion of crossterms on autoterm regions cannot be prevented or entirely removed. This is because the window spreads out the crossterms in the timefrequency domain so that the mainlobe or/and the sidelobes of the crossterms are deemed to overlap with the signal autoterms. We refer to this undesired property as the leakage problem in STFDs. In addition to the above drawbacks, the timefrequency kernel ignores the fact that the first type of crossterms need not be smoothed, as its appearance along the diagonal elements can improve the effective signal-to-noise ratio. The spatial averaging of the STFD defined in (23) at a given point does not smooth or reduce the crossterms at that point but rather moves them from their residence on the off-diagonal matrix entries to be part of the matrix diagonal elements. The other part represents the contribution of the autoterms at the same point. Therefore, not only we are able to set the off-diagonal elements of the source TFD matrix to zeros, but we can points of peak also improve performance by selecting the values, irrespective of whether these points belong to autoterm or crossterm regions. IV. PERFORMANCE EVALUATION A. Performance Index We use a slightly modified version of the performance index applied in [2] to evaluate the performance of the proposed source separation technique. The estimate of the source signal vector is computed by applying the pseudo-inverse of the

estimated mixing matrix i.e.,

to the received signal vector

(26) . We stress that in general, this procedure is where not optimal for recovering the source signals based on an estimate . For large enough sample size, matrix should be close well approximates the identity to the true one so that matrix. We normalize matrix by (27) denotes the matrix formed by the diagonal where become elements of . As such, the diagonal elements of exactly one, giving more meaning to the performance index (28) which defines the interference-to-signal ratio (ISR). Thus, measures the ratio of the power of the interference of th source signal to the power of the th source signal. As the measure of the global quality of the separation process, we also apply the global rejection level to evaluate the overall performance of the proposed method (29)

B. Effect of Crossterms between Source Signals In this section, we examine the effect of the timefrequency crossterms on source separation performance when spatial averaging is not applied. To simplify the problem, we assume that is an identity matrix. When crossterms are present at the , then off-diagonal elements of the TFD matrix (30) is the diagonal matrix with the eigenvalues at the where is the matrix whose columns are diagonal elements, and the corresponding eigenvectors. Note that all the above matrices point. Substituting (30) in (8), depend on the selected the STFD matrix of the data vector under noise-free conditions becomes

(31) and the STFD matrix of the whitened array signal vector is

(32) is diagonal, is unitary. If the estiSince is provided based on a single mated mixing matrix point, then from (32) (33)

ZHANG AND AMIN: SPATIAL AVERAGING OF TIMEFREQUENCY DISTRIBUTIONS FOR SIGNAL RECOVERY

2897

which is dependent on the unitary matrix

. Furthermore (34)

and

..

Fig. 2.

TFD of the source signals (WignerVille distribution).

. . .

. . .

..

. . . C. Simulation Results In this section, we demonstrate the effectiveness of the spatially averaged timefrequency distributions in source separations. The whitening joint-diagonalization scheme [2] is used for incorporating multiple timefrequency points into the proposed spatial averaging method. In all simulations, two sources with the chirp signals (35) (38) . Different values of are used, where is chosen as and are considered. These values control the frequency offset and amplitude variation between the two signals and can be chosen to yield closely or widely separated source signatures in the timefrequency domain. We consider 128 data samples is formed. from which a timefrequency matrix of and are set equal to The DOAs of the two signals and , respectively, from the broadside direction. Furthermore, we assume an equispaced five-element linear array , where is the wavelength. with the interelement spacing Subsequently, when the spatial averaging method is used, two subarrays are formed, each with three elements. , i.e., neither In the first set of simulations, we choose signal is amplitude modulated. The WignerVille (WV) distribution of each signal is shown in Fig. 2, where is set equal to 0.05. Fig. 3 shows the timefrequency distribution of the mixed signals at the center array sensor. No noise is present for this case. It is clear that the crossterms lie in the middle of the two chirps, and their amplitude changes periodically. Fig. 4(a) shows the timefrequency distributions of the separated signals using the technique in [2], where joint diagonalization is used without the utilization of the proposed spatial points are used averaging method. Three timefrequency 32, 64, and 96. The frequency is chosen so that the at TFD computed using the data at the center array sensor is the largest at each . Peak values of the WV distribution may either correspond to autoterms or crossterms. In this case, out of three

. . .

. . .

..

. . .

where becomes

. Accordingly, the performance index

(36) and the global rejection level is given by

(37) are always In general, since the absolute values of takes a equal to or smaller than 1, the global rejection level only when positive value. It is clear that holds true for all . That is, is an identity matrix, which implies that there is no off-diagonal nonzero elements in matrix . . If we select a point Consider the specific case of where the contributions of the two sources to the source TFD , and since matrix are the same, i.e., by definition, then it is straightfor. In this case, is conward to show that points having such property include stant equal to 2. The all crossterms at which the autoterms have equal contributions.

2898

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 48, NO. 10, OCTOBER 2000

points, one crossterm point and two autoterm peaks were matrix is selected. The obtained

Fig. 3.

TFD of the mixed signal (WignerVille distribution).

and the computed global rejection level is 0.43 dB. is far The result is clearly unsatisfactory, as the matrix from the identity matrix, and crossterms appear in the separated signals. Next, we force the selection of autoterm peaks by only conpoints along the instantaneous frequencies of sidering the the two input signals at the same above time instants. Although no crossterm point is selected, yet as discussed in Section III, because of the finite data record, the crossterms leak into autoterm regions, causing the source TFD matrix to deviate from a diagonal structure. We show in Fig. 4(b) the result of source separation when only the autoterm points are considered. The matrix becomes obtained

(a)

(b) Fig. 4. TFD of the separated signals without spatial averaging (WignerVille distribution). (a) Using peak timefrequency points. (b) Using autoterm points.

and the respective global rejection level is 23.96 dB. It is clear that the source separation performance is greatly improved. This good performance implies that the contributions of crossterms at the three selected autoterm points were insignificant, implying that the corresponding source TFD matrices in this case were close to diagonal. Fig. 5 shows the timefrequency distributions of the separated signals at the same condition as Fig. 4(a), except with the utilization of the proposed spatial averaging method. Spatial averaging entirely mitigates the effect of crossterms. It is clear that the timefrequency distributions of the separated signals are the are exsame as those of the original source signals, and actly identity matrices. Similar results can be obtained when all points are autoterms. three versus the freFig. 6 shows the global rejection level between the two chirps, where the input quency difference SNR is 20 dB. When the proposed spatial averaging method is used, the global rejection level maintains low values. On the other hand, without spatial averaging, the global rejection levels become very high. The main reason of the large fluctuation of without spatial averaging is that the influence as well the as the number of crossterm points incorporated in the joint-diagonalization scheme varies with the frequency difference (when , no crossterm points were selected). Note that the crossterms of the WignerVille distribution remain high even when the frequency difference is large. When selected, these terms put large values along the off-diagonal terms of the source TFD matrix and subsequently cause considerable error, as is evident from the figure. However, when only autoterm points are used, the global rejection level decreases as increases. In this case, the matrix off-diagonal elements are the crossterm values at the autoterm points that become smaller for higher values of . Next, we show the effect of using timefrequency smoothing kernels for reduced interference terms. The ChoiWilliams

ZHANG AND AMIN: SPATIAL AVERAGING OF TIMEFREQUENCY DISTRIBUTIONS FOR SIGNAL RECOVERY

2899

Fig. 5. TFD of the separated signals with spatial averaging (WignerVille distribution).

Fig. 7. TFD of the source signals (ChoiWilliams distribution).

Fig. 8. TFD of the mixed signal (ChoiWilliams distribution). Fig. 6. Global rejection level versus frequency difference f (WignerVille distribution). (Input SNR 20 dB; o: without spatial averaging; : with spatial averaging; : without spatial averaging using autoterm points; : with spatial averaging using autoterm points).

4 5

three used, the

points. When only the autoterm matrix becomes

points are

(CW) distribution [10] is considered with . Fig. 7 shows the CW distribution of each signal separately, whereas the CW distribution of the mixed signals at the center array sensor is depicted in Fig. 8. The signals are the same as the ones used . Fig. 9(a) in the WV distribution simulations with shows the CW distributions of the separated signals. The matrix is obtained

and the respective global rejection level is 0.26 dB. At this small frequency offset, effective smoothing of crossterms is difficult, and as a result, even with the use of timefrequency point was still selected out of the kernel, one crossterm

and the global rejection level is reduced to 12.86 dB. The CW distributions of the separated signals are shown in Fig. 9(b). Fig. 10 shows the CW distributions of the separated signals under the same condition as Fig. 9(a), with the utilization of the spatial averaging method. Again, it is clear that the timefrequency distributions of both cases are the same as the source are exactly an identity matrix. The same resignals, and points are sults can be obtained when only the autoterm used. Fig. 11 shows the global rejection level versus the frequency between the two chirps, where the input SNR difference is 20 dB. It is evident from this figure that the kernel method fails when the two signals have close timefrequency signatures.

2900

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 48, NO. 10, OCTOBER 2000

(a)

Fig. 10. TFD of the separated signals with spatial averaging (ChoiWilliams distribution).

(b) Fig. 9. TFD of the separated signals without spatial averaging (ChoiWilliams distribution). (a) Using peak timefrequency points. (b) Using autoterm points. Fig. 11. Global rejection level versus frequency difference f (ChoiWilliams distribution). (Input SNR 20 dB; o: without spatial averaging; : with spatial averaging; : without spatial averaging using autoterm points; : with spatial averaging using autoterm points).

Using the proposed spatial averaging method outperforms the case when no spatial averaging is applied. Three important observations on the difference between the WV distribution and the CW distribution in the context of source separation are in order. First, the CW kernel effectively reduces the crossterms, is large. Accordingly, crossterms are not particularly when as large as the autoterms, and as such, it is unlikely for the crossterms to be selected and incorporated in the joint-diagois large enough, the global nalization scheme. Second, when rejection level is significantly reduced for the CW distribution, even when spatial averaging is not applied. Third, when the spatial averaging method is used, the performance at small frequency offset from the CW distribution is worse than that obtained from the WV distribution. The reason is that source separation is perturbed by the presence of noise, and the performance nevertheless is sensitive to the input SNR. When comparing the WV distribution and the CW distribution, the noise floor relative to peak values is lower in the WV distribution than in the CW for the underlying chirp signal example.

4 5

To show the effect of the input SNR on the source separation performance, Figs. 12 and 13 depict the global rejection level versus the input SNR, where the frequency difference is 0.01. Increasing the SNR certainly improves the source separation performance when spatial averaging is applied. On the other hand, without spatial averaging, the source separation performance holds an almost constant high level. Such a performance demonstrates that crossterms are more of a fundamental problem than noise in TFD-based source-separation problems. In the second set of simulations, is set to zero in (38), rendering the two source signals identical in terms of their instantaneous frequency characteristic. However, one of the two source signals is amplitude modulated, which is caused by a nonzero positive value of . Fig. 14 shows the global rejection level versus , where the WV distribution is considered, and the input SNR is 20 dB. It is clear that the two signals cannot be separated without spatial

ZHANG AND AMIN: SPATIAL AVERAGING OF TIMEFREQUENCY DISTRIBUTIONS FOR SIGNAL RECOVERY

2901

Fig. 12. Global rejection level versus input SNR (WignerVille distribution). 0:01; o: without spatial averaging; : with spatial averaging; : (f without spatial averaging using autoterm points; : with spatial averaging using autoterm points).

Fig. 14. Global rejection level versus  (WignerVille distribution). (Input SNR = 20 dB; o: without spatial averaging; : with spatial averaging).

Fig. 13. Global rejection level versus input SNR (ChoiWilliams distribution). (f = 0:01; o: without spatial averaging; : with spatial averaging; : without spatial averaging using autoterm points; : with spatial averaging using autoterm points).

averaging, but when applying spatial averaging, satisfactory per, formance of source separation can be achieved. For the proposed technique yields a global rejection level 26.72 dB. V. CONCLUSIONS Spatial averaging of spatial timefrequency distributions has been introduced and the role of spatial averaging in mitigating the effects of crossterms when bilinear transforms are used for signal recovery has been shown. The spatial averaging of the spatial timefrequency distributions of the data across an antenna array removes the undesired effect of crossterms between the impinging signals. These terms reside along the off-diagonal entries of the source timefrequency distribution matrix and consequently impede the source separation performance,

which is based on preassumed diagonal matrix structure. Spatial averaging amounts to forming a spatial Hermitian Toeplitz matrix using the auto- and cross-timefrequency distributions of the data over one half of the uniform linear array. This matrix is then added to the spatial matrix corresponding to the other half of the array. The desired effect of this averaging is reallocating the interaction between the source signals in the timefrequency domain from the off-diagonal to the diagonal elements of the source TFD matrix. In this respect, unlike the method proposed in [2], cross-terms, due to their potential high values, are regarded as useful components that could be used for improved performance. Spatial averaging can be applied to all members of Cohens class of TFDs, irrespective of the employed smoothing kernel. When using a timefrequency kernel, the problem amounts to averaging in all three dimensions of time, frequency, and space. Joint-diagonalization (JD) is applied to include multiple spatially averaged timefrequency distributions at different timefrequency points. With cross-terms moved to the diagonal entries of the TFD matrix, the prime task of the source separation based on the JD scheme is to avoid degenerate eigenvalues that are responsible for the nonuniqueness solution of the problem. Simulation examples were presented to illustrate the effectiveness of the new approach. The two performance measures used were the global rejection level and the values of the off-diagonal elements of the product of the mixing matrix and the Pseudo inverse of its estimate. Two sources and five sensors were considered. The source signals were chirp signals with the same sweeping frequency, but their corresponding constant frequencies and amplitudes were offset by different values. Both WignerVille and ChoiWilliams distributions were considered. It was shown that the spatial averaging method significantly improves the performance measures over the nonspatially averaging method, specifically when the two signals have close timefrequency signatures. Without spatial averaging, performance is very sensitive to whether only auto-term or cross-term points or their mix are

2902

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 48, NO. 10, OCTOBER 2000

incorporated in the source separation procedure. With spatial averaging, this is no longer a concern since both terms appear along the diagonal. It is also shown that the ChoiWilliams distribution provides better results than the WignerVille distribution when no spatial averaging is applied since it lowers the likelihood of selecting crossterm points. With spatial averaging, the issue becomes merely SNR, and in this respect, the WignerVille distribution, due to its high peak values, yields better performance than the ChoiWilliams distribution. Therefore, the timefrequency smoothing becomes unnecessary whenever spatial array averaging is possible. REFERENCES
[1] A. Belouchrani and M. Amin, Source separation based on the diagonalization of a combined set of spatial timefrequency distribution matrices, in Proc. ICASSP, Munich, Germany, Apr. 1997. , Blind source separation based on timefrequency signal repre[2] sentation, IEEE Trans. Signal Processing, vol. 46, pp. 28882898, Nov. 1998. , Blind source separation using joint signal representations, in [3] Proc. SPIE: Advanced Algorithms Architectures Signal Process., San Diego, CA, Aug. 1997. [4] M. G. Amin and A. Belouchrani, Blind source separation using the spatial ambiguity functions, in Proc. IEEE Int. Symp. TimeFreq. TimeScale Anal., Pittsburgh, PA, Oct. 1998, pp. 413416. [5] A. Belouchrani and M. Amin, Timefrequency MUSIC, IEEE Signal Processing Lett., vol. 6, pp. 109110, May 1999. [6] S. U. Pillai, Array Signal Processing. New York: Springer-Verlag, 1989. [7] L. Cohen, TimeFrequency Analysis. Englewood Cliffs, NJ: PrenticeHall, 1995. [8] L. Tong, Y. Inouye, and R.-W. Liu, Waveform-preserving blind estimation of multiple independent sources, IEEE Trans. Signal Processing, vol. 41, pp. 24612470, July 1993. [9] A. Belouchrani, K. A. Meraim, J.-F. Cardoso, and E. Moulines, A blind source separation techniques using second order statistics, IEEE Trans. Signal Processing, vol. 45, pp. 434444, Feb. 1997. [10] H. I. Choi and W. J. Williams, Improved timefrequency representation of multicomponent signals using exponential kernels, IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp. 862871, June 1989. [11] P. Stoica and K. Sharman, Maximum likelihood methods for direction-of-arrival estimation, IEEE Trans. Acoust., Speech, Signal Processing, vol. 38, pp. 11321143, July 1990. [12] Y. Hua and T. K. Sarkar, Matrix pencil method for estimating parameters of exponentially damped/undamped sinusoids in noise, IEEE Trans. Acoust., Speech, Signal Processing, vol. 38, pp. 814824, May 1990. [13] R. O. Schmidt, Multiple emitter location and signal parameter estimation, IEEE Trans. Antennas Propagat., vol. 34, pp. 276280, Mar. 1986. [14] A. J. Barabell, Improving the resolution performance of eigenstructure-based direction-finding algorithms, in Proc. ICASSP, Boston, MA, 1983, pp. 336339. [15] B. D. Rao and K. V. S. Hari, Performance analysis of root-MUSIC, IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp. 19391949, Dec. 1989. [16] R. Roy and T. Kailath, ESPRITEstimation of signal parameters via rotational invariance techniques, IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp. 984995, July 1989. [17] Y. Zhang, W. Mu, and M. G. Amin, Subspace analysis of spatial timefrequency distribution matrices, IEEE Trans. Signal Processing, submitted for publication.

[18] Y. Zhang, W. Mu, and M. G. Amin, Timefrequency maximum likelihood methods for direction finding, J. Franklin Inst., vol. 337, pp. 483497, July 2000. [19] Y. Zhang and M. G. Amin, Blind separation of sources based on their timefrequency signatures, in Proc. ICASSP, Istanbul, Turkey, June 2000. [20] J. E. Evans, J. R. Johnson, and D. F. Sun, High resolution angular spectrum estimation techniques for terrain scattering analysis and angle of arrival estimation, in Proc. First Acoust., Speech, Signal Process. Workshop Spectral Estim., Hamilton, ON, Canada, Aug. 1981. [21] T. J. Shan and T. Kailath, On spatial smoothing for DOA estimation of coherent sources, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-33, pp. 806811, 1985. [22] R. T. Williams, S. Prasad, A. K. Mahalanabis, and L. H. Sibul, An improved spatial smoothing technique for bearing estimation in multipath environment, IEEE Trans. Acoust., Speech, Signal Processing, vol. 36, pp. 425432, 1988. [23] S. U. Pillai and B. H. Kwon, Forward/backward spatial smoothing techniques for coherent signal identification, IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp. 815, Jan. 1989. [24] S. U. Pillai, Y. Barness, and F. Haber, A new approach to array geometry for improved spatial spectrum estimation, Proc. IEEE, vol. 73, pp. 15221524, Oct. 1985.

Yimin Zhang (M88) received the M.S. and Ph.D. degrees from the University of Tsukuba, Tsukuba, Japan, in 1985 and 1988, respectively. He joined the faculty of the Department of Radio Engineering, Southeast University, Shanghai, China, in 1988. He served as a Technical Manager at Communication Laboratory Japan from 1995 to 1997 and a Visiting Researcher at ATR Adaptive Communications Research Laboratories, Tokyo, Japan, from 1997 to 1998. Currently, he is a Research Fellow with the Department of Electrical and Computer Engineering, Villanova University, Villanova, PA. His current research interests are in the areas of array signal processing, space-time adaptive processing, blind signal processing, digital mobile communications, and timefrequency analysis.

Moeness G. Amin (SM91) received the Ph.D. degree in electrical engineering in 1984 from University of Colorado, Boulder. He has been on the Faculty of the Department of Electrical and Computer Engineering at Villanova University, Villanova, PA, since 1985, where is now a Professor. His current research interests are in the areas of timefrequency analysis, spread spectrum communications, smart antennas, and blind signal processing. Dr. Amin was an Associate Editor of the IEEE TRANSACTIONS ON SIGNAL PROCESSING from 1995 to 1997 and a Member of the Technical Committee of the IEEE Signal Processing Technical Committee on Signal Processing for Communications. He was the General Chair of the 1994 IEEE International Symposium on TimeFrequency and TimeScale Analysis. He is the General Chair of the 2000-IEEE Workshop on Statistical Signal and Array Processing. He was the recipient of the 1997 IEEE Philadelphia Section Service Award and the IEEE Third Millennium Medal. He was also the recipient of the 1997 Villanova University Outstanding Faculty Research Award. He serves on the Committee of Arts and Science of the Franklin Institute.